data-archetype commited on
Commit
969b5a3
·
0 Parent(s):

MOSS-Image-Base 8-step DMD2 LoRA

Browse files
.gitattributes ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tar filter=lfs diff=lfs merge=lfs -text
29
+ *.tflite filter=lfs diff=lfs merge=lfs -text
30
+ *.tgz filter=lfs diff=lfs merge=lfs -text
31
+ *.wasm filter=lfs diff=lfs merge=lfs -text
32
+ *.xz filter=lfs diff=lfs merge=lfs -text
33
+ *.zip filter=lfs diff=lfs merge=lfs -text
34
+ *.zst filter=lfs diff=lfs merge=lfs -text
35
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ images/example_goose.png filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ Apache License
3
+ Version 2.0, January 2004
4
+ http://www.apache.org/licenses/
5
+
6
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
7
+
8
+ 1. Definitions.
9
+
10
+ "License" shall mean the terms and conditions for use, reproduction,
11
+ and distribution as defined by Sections 1 through 9 of this document.
12
+
13
+ "Licensor" shall mean the copyright owner or entity authorized by
14
+ the copyright owner that is granting the License.
15
+
16
+ "Legal Entity" shall mean the union of the acting entity and all
17
+ other entities that control, are controlled by, or are under common
18
+ control with that entity. For the purposes of this definition,
19
+ "control" means (i) the power, direct or indirect, to cause the
20
+ direction or management of such entity, whether by contract or
21
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
22
+ outstanding shares, or (iii) beneficial ownership of such entity.
23
+
24
+ "You" (or "Your") shall mean an individual or Legal Entity
25
+ exercising permissions granted by this License.
26
+
27
+ "Source" form shall mean the preferred form for making modifications,
28
+ including but not limited to software source code, documentation
29
+ source, and configuration files.
30
+
31
+ "Object" form shall mean any form resulting from mechanical
32
+ transformation or translation of a Source form, including but
33
+ not limited to compiled object code, generated documentation,
34
+ and conversions to other media types.
35
+
36
+ "Work" shall mean the work of authorship, whether in Source or
37
+ Object form, made available under the License, as indicated by a
38
+ copyright notice that is included in or attached to the work
39
+ (an example is provided in the Appendix below).
40
+
41
+ "Derivative Works" shall mean any work, whether in Source or Object
42
+ form, that is based on (or derived from) the Work and for which the
43
+ editorial revisions, annotations, elaborations, or other modifications
44
+ represent, as a whole, an original work of authorship. For the purposes
45
+ of this License, Derivative Works shall not include works that remain
46
+ separable from, or merely link (or bind by name) to the interfaces of,
47
+ the Work and Derivative Works thereof.
48
+
49
+ "Contribution" shall mean any work of authorship, including
50
+ the original version of the Work and any modifications or additions
51
+ to that Work or Derivative Works thereof, that is intentionally
52
+ submitted to Licensor for inclusion in the Work by the copyright owner
53
+ or by an individual or Legal Entity authorized to submit on behalf of
54
+ the copyright owner. For the purposes of this definition, "submitted"
55
+ means any form of electronic, verbal, or written communication sent
56
+ to the Licensor or its representatives, including but not limited to
57
+ communication on electronic mailing lists, source code control systems,
58
+ and issue tracking systems that are managed by, or on behalf of, the
59
+ Licensor for the purpose of discussing and improving the Work, but
60
+ excluding communication that is conspicuously marked or otherwise
61
+ designated in writing by the copyright owner as "Not a Contribution."
62
+
63
+ "Contributor" shall mean Licensor and any individual or Legal Entity
64
+ on behalf of whom a Contribution has been received by Licensor and
65
+ subsequently incorporated within the Work.
66
+
67
+ 2. Grant of Copyright License. Subject to the terms and conditions of
68
+ this License, each Contributor hereby grants to You a perpetual,
69
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
70
+ copyright license to reproduce, prepare Derivative Works of,
71
+ publicly display, publicly perform, sublicense, and distribute the
72
+ Work and such Derivative Works in Source or Object form.
73
+
74
+ 3. Grant of Patent License. Subject to the terms and conditions of
75
+ this License, each Contributor hereby grants to You a perpetual,
76
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
77
+ (except as stated in this section) patent license to make, have made,
78
+ use, offer to sell, sell, import, and otherwise transfer the Work,
79
+ where such license applies only to those patent claims licensable
80
+ by such Contributor that are necessarily infringed by their
81
+ Contribution(s) alone or by combination of their Contribution(s)
82
+ with the Work to which such Contribution(s) was submitted. If You
83
+ institute patent litigation against any entity (including a
84
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
85
+ or a Contribution incorporated within the Work constitutes direct
86
+ or contributory patent infringement, then any patent licenses
87
+ granted to You under this License for that Work shall terminate
88
+ as of the date such litigation is filed.
89
+
90
+ 4. Redistribution. You may reproduce and distribute copies of the
91
+ Work or Derivative Works thereof in any medium, with or without
92
+ modifications, and in Source or Object form, provided that You
93
+ meet the following conditions:
94
+
95
+ (a) You must give any other recipients of the Work or
96
+ Derivative Works a copy of this License; and
97
+
98
+ (b) You must cause any modified files to carry prominent notices
99
+ stating that You changed the files; and
100
+
101
+ (c) You must retain, in the Source form of any Derivative Works
102
+ that You distribute, all copyright, patent, trademark, and
103
+ attribution notices from the Source form of the Work,
104
+ excluding those notices that do not pertain to any part of
105
+ the Derivative Works; and
106
+
107
+ (d) If the Work includes a "NOTICE" text file as part of its
108
+ distribution, then any Derivative Works that You distribute must
109
+ include a readable copy of the attribution notices contained
110
+ within such NOTICE file, excluding those notices that do not
111
+ pertain to any part of the Derivative Works, in at least one
112
+ of the following places: within a NOTICE text file distributed
113
+ as part of the Derivative Works; within the Source form or
114
+ documentation, if provided along with the Derivative Works; or,
115
+ within a display generated by the Derivative Works, if and
116
+ wherever such third-party notices normally appear. The contents
117
+ of the NOTICE file are for informational purposes only and
118
+ do not modify the License. You may add Your own attribution
119
+ notices within Derivative Works that You distribute, alongside
120
+ or as an addendum to the NOTICE text from the Work, provided
121
+ that such additional attribution notices cannot be construed
122
+ as modifying the License.
123
+
124
+ You may add Your own copyright statement to Your modifications and
125
+ may provide additional or different license terms and conditions
126
+ for use, reproduction, or distribution of Your modifications, or
127
+ for any such Derivative Works as a whole, provided Your use,
128
+ reproduction, and distribution of the Work otherwise complies with
129
+ the conditions stated in this License.
130
+
131
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
132
+ any Contribution intentionally submitted for inclusion in the Work
133
+ by You to the Licensor shall be under the terms and conditions of
134
+ this License, without any additional terms or conditions.
135
+ Notwithstanding the above, nothing herein shall supersede or modify
136
+ the terms of any separate license agreement you may have executed
137
+ with Licensor regarding such Contributions.
138
+
139
+ 6. Trademarks. This License does not grant permission to use the trade
140
+ names, trademarks, service marks, or product names of the Licensor,
141
+ except as required for reasonable and customary use in describing the
142
+ origin of the Work and reproducing the content of the NOTICE file.
143
+
144
+ 7. Disclaimer of Warranty. Unless required by applicable law or
145
+ agreed to in writing, Licensor provides the Work (and each
146
+ Contributor provides its Contributions) on an "AS IS" BASIS,
147
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
148
+ implied, including, without limitation, any warranties or conditions
149
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
150
+ PARTICULAR PURPOSE. You are solely responsible for determining the
151
+ appropriateness of using or redistributing the Work and assume any
152
+ risks associated with Your exercise of permissions under this License.
153
+
154
+ 8. Limitation of Liability. In no event and under no legal theory,
155
+ whether in tort (including negligence), contract, or otherwise,
156
+ unless required by applicable law (such as deliberate and grossly
157
+ negligent acts) or agreed to in writing, shall any Contributor be
158
+ liable to You for damages, including any direct, indirect, special,
159
+ incidental, or consequential damages of any character arising as a
160
+ result of this License or out of the use or inability to use the
161
+ Work (including but not limited to damages for loss of goodwill,
162
+ work stoppage, computer failure or malfunction, or any and all
163
+ other commercial damages or losses), even if such Contributor
164
+ has been advised of the possibility of such damages.
165
+
166
+ 9. Accepting Warranty or Additional Liability. While redistributing
167
+ the Work or Derivative Works thereof, You may choose to offer,
168
+ and charge a fee for, acceptance of support, warranty, indemnity,
169
+ or other liability obligations and/or rights consistent with this
170
+ License. However, in accepting such obligations, You may act only
171
+ on Your own behalf and on Your sole responsibility, not on behalf
172
+ of any other Contributor, and only if You agree to indemnify,
173
+ defend, and hold each Contributor harmless for any liability
174
+ incurred by, or claims asserted against, such Contributor by reason
175
+ of your accepting any such warranty or additional liability.
176
+
177
+ END OF TERMS AND CONDITIONS
178
+
179
+ APPENDIX: How to apply the Apache License to your work.
180
+
181
+ To apply the Apache License to your work, attach the following
182
+ boilerplate notice, with the fields enclosed by brackets "[]"
183
+ replaced with your own identifying information. (Don't include
184
+ the brackets!) The text should be enclosed in the appropriate
185
+ comment syntax for the file format. We also recommend that a
186
+ file or class name and description of purpose be included on the
187
+ same "printed page" as the copyright notice for easier
188
+ identification within third-party archives.
189
+
190
+ Copyright [yyyy] [name of copyright owner]
191
+
192
+ Licensed under the Apache License, Version 2.0 (the "License");
193
+ you may not use this file except in compliance with the License.
194
+ You may obtain a copy of the License at
195
+
196
+ http://www.apache.org/licenses/LICENSE-2.0
197
+
198
+ Unless required by applicable law or agreed to in writing, software
199
+ distributed under the License is distributed on an "AS IS" BASIS,
200
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
201
+ See the License for the specific language governing permissions and
202
+ limitations under the License.
README.md ADDED
@@ -0,0 +1,163 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: OpenMOSS-Team/MOSS-Image-Base
3
+ base_model_relation: adapter
4
+ language:
5
+ - en
6
+ - zh
7
+ license: apache-2.0
8
+ library_name: peft
9
+ pipeline_tag: text-to-image
10
+ tags:
11
+ - moss-image
12
+ - diffusers
13
+ - lora
14
+ - text-to-image
15
+ - distillation
16
+ ---
17
+ # moss_image_base_dmd2
18
+
19
+ An 8-step LoRA (rank 128) for [MOSS-Image-Base](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base), distilled by distribution
20
+ matching: 8 forwards at CFG 1, without a negative prompt, in place of the base model's
21
+ 50 guided steps. It adapts only the transformer; the text encoder (Qwen3.5-9B), VAE
22
+ (FLUX.2 small decoder), tokenizer and inference pipeline are MOSS-Image-Base's,
23
+ unchanged. How it was trained is in [TECH_REPORT.md](TECH_REPORT.md).
24
+
25
+ It was trained on MOSS-Image-Base, but should work with most fine-tunes of it and with
26
+ other LoRAs: load it alongside them, at strength 1.
27
+
28
+ ![A goose resting by a rocky pond, generated in 8 steps](images/example_goose.png)
29
+
30
+ *8 steps at CFG 1, 1152 × 896, seed 42, with the prompt in the Diffusers example below. The image carries its ComfyUI workflow: drag it into ComfyUI to load it.*
31
+
32
+ | File | Format |
33
+ | --- | --- |
34
+ | `loras/moss_image_base_dmd2_bf16.safetensors` | ComfyUI LoRA (kohya keys on ComfyUI's module names, q/k/v as one LoRA on the fused `qkv`) |
35
+ | `transformer/` | PEFT LoRA adapter for the Diffusers `MOSSImageTransformer` |
36
+ | `images/example_goose.png` | Example image, its ComfyUI workflow embedded |
37
+ | `workflows/moss_image_base_dmd2_8step.json` | ComfyUI workflow (drag into ComfyUI) |
38
+ | `workflows/moss_image_base_dmd2_8step_api.json` | The same workflow in ComfyUI's API format |
39
+
40
+ Both LoRA files make the same weight change; the ComfyUI file packs each block's q, k and v
41
+ LoRAs into one exactly equivalent LoRA on ComfyUI's fused `qkv` projection.
42
+
43
+ ## Sampling
44
+
45
+ The LoRA is trained for one sampler; others give worse results.
46
+
47
+ | Setting | Value |
48
+ | --- | --- |
49
+ | Steps | 8 |
50
+ | Sampler | At each step, take the model's clean estimate, then re-noise it with fresh noise to the next level (ComfyUI's `lcm`) |
51
+ | Schedule | 8 evenly spaced times 1, 7/8, ..., 1/8, then 0, under the flow shift 4.4817 (ComfyUI's `simple` scheduler after a ModelSamplingAuraFlow node at shift 4.4817); about 1, .969, .931, .882, .818, .729, .599, .39 |
52
+ | CFG | 1, no negative prompt (guidance is distilled in) |
53
+ | LoRA strength | 1 |
54
+ | Resolution | About 1 megapixel, any aspect ratio the base supports |
55
+
56
+ ## ComfyUI
57
+
58
+ Use ComfyUI's native MOSS-Image support and load the workflow: drag `images/example_goose.png` or
59
+ `workflows/moss_image_base_dmd2_8step.json` into ComfyUI.
60
+
61
+ | File | Put in | Loader |
62
+ | --- | --- | --- |
63
+ | `loras/moss_image_base_dmd2_bf16.safetensors` (this repo) | `models/loras/` | LoraLoaderBypassModelOnly |
64
+ | `diffusion_models/moss_image_base_bf16.safetensors`, from Comfy-Org/MOSS-Image | `models/diffusion_models/` | Load Diffusion Model |
65
+ | `text_encoders/qwen3.5_9b_bf16.safetensors` (or its int8 / w4a8 variants), from Comfy-Org/MOSS-Image | `models/text_encoders/` | Load CLIP, type `moss_image` |
66
+ | `vae/flux2-vae_small.safetensors`, from Comfy-Org/MOSS-Image | `models/vae/` | Load VAE |
67
+
68
+ The workflow sets the sampler (`lcm`), scheduler (`simple`), 8 steps, CFG 1 and the
69
+ shift. ComfyUI's default MOSS-Image shift is 5, close to but not the 4.4817 the LoRA
70
+ was trained with; the workflow's ModelSamplingAuraFlow node sets it. Use that node (or
71
+ ModelSamplingFlux) to change the shift, not ModelSamplingSD3: it scales MOSS-Image's
72
+ timesteps differently and the output turns to noise.
73
+
74
+ Load the LoRA with the bypass loader ("Load LoRA (Bypass, Model Only)"), as the workflow does,
75
+ not the standard LoRA loader. The standard loader merges the LoRA into the bf16 weights, and
76
+ the weight change is smaller than bf16's rounding step, so most of it is lost. The bypass
77
+ loader applies it unmerged, as Diffusers does below.
78
+
79
+ ## Diffusers
80
+
81
+ Install a PyTorch build that matches your CUDA driver first
82
+ ([pytorch.org](https://pytorch.org/get-started/locally/)), then:
83
+
84
+ ```bash
85
+ pip install "diffusers>=0.38" "transformers>=5.3" "accelerate>=1.10" "safetensors>=0.6" "peft>=0.21" "Pillow>=10"
86
+ ```
87
+
88
+ ```python
89
+ import math
90
+
91
+ import torch
92
+ from diffusers import AutoencoderKLFlux2, FlowMatchEulerDiscreteScheduler
93
+ from diffusers.utils.dynamic_modules_utils import get_class_from_dynamic_module
94
+ from peft import PeftModel
95
+ from transformers import AutoModel, AutoTokenizer
96
+
97
+ BASE = "OpenMOSS-Team/MOSS-Image-Base"
98
+ SHIFT = math.exp(1.5)
99
+
100
+
101
+ class FreshNoiseScheduler(FlowMatchEulerDiscreteScheduler):
102
+ """The LoRA's sampler: jump to the clean estimate, re-noise with fresh noise."""
103
+
104
+ def set_timesteps(self, num_inference_steps=8, device=None):
105
+ t = torch.linspace(1.0, 1.0 / num_inference_steps, num_inference_steps)
106
+ sigmas = SHIFT * t / (1 + (SHIFT - 1) * t)
107
+ super().set_timesteps(device=device, sigmas=sigmas.tolist())
108
+
109
+ def step(self, model_output, timestep, sample, return_dict=True, **kwargs):
110
+ if self.step_index is None:
111
+ self._init_step_index(timestep)
112
+ sigma, sigma_next = self.sigmas[self.step_index], self.sigmas[self.step_index + 1]
113
+ clean = sample.float() - sigma * model_output.float()
114
+ if sigma_next > 0:
115
+ clean = (1 - sigma_next) * clean + sigma_next * torch.randn_like(clean)
116
+ self._step_index += 1
117
+ return (clean.to(sample.dtype),)
118
+
119
+
120
+ def moss_class(name):
121
+ return get_class_from_dynamic_module(
122
+ BASE, "pipeline_moss_image.py", class_name=name, trust_remote_code=True
123
+ )
124
+
125
+
126
+ transformer = moss_class("MOSSImageTransformer").from_pretrained(
127
+ "OpenMOSS-Team/MOSS-Image-Base", subfolder="transformer", torch_dtype=torch.bfloat16
128
+ )
129
+ transformer = PeftModel.from_pretrained(
130
+ transformer, "data-archetype/moss_image_base_dmd2", subfolder="transformer"
131
+ ).get_base_model()
132
+
133
+ pipe = moss_class("MOSSImagePipeline")(
134
+ transformer=transformer,
135
+ scheduler=FreshNoiseScheduler(shift=1.0),
136
+ text_encoder=AutoModel.from_pretrained(BASE, subfolder="text_encoder", dtype=torch.bfloat16),
137
+ tokenizer=AutoTokenizer.from_pretrained(BASE, subfolder="tokenizer"),
138
+ vae=AutoencoderKLFlux2.from_pretrained(BASE, subfolder="vae", torch_dtype=torch.float32),
139
+ )
140
+ pipe.to("cuda") # or pipe.enable_model_cpu_offload() on a smaller GPU
141
+
142
+ torch.manual_seed(42)
143
+ image = pipe(
144
+ "A large brown and white goose resting on dry grass at the edge of a rocky pond, its patterned wing feathers folded neatly and its long neck curved as it looks toward the water. The sun casts warm light across the bird and the rough stone bank behind it.",
145
+ height=896,
146
+ width=1152,
147
+ num_inference_steps=8,
148
+ guidance_scale=1.0,
149
+ uncond_skip_layers=None,
150
+ ).images[0]
151
+ ```
152
+
153
+ The adapter is applied unmerged, at strength 1. Prompt formats (natural language, tags
154
+ and JSON layouts) and requirements are the base model's; see MOSS-Image-Base's
155
+ [model card](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base) and
156
+ [user guide](https://github.com/OpenMOSS/MOSS-Image/blob/main/USERGUIDE.md).
157
+
158
+ ## License
159
+
160
+ The weights are released under the base model's license, Apache 2.0
161
+ ([LICENSE](LICENSE), copied from OpenMOSS-Team/MOSS-Image-Base).
162
+
163
+ We thank the OpenMOSS team for MOSS-Image.
TECH_REPORT.md ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # moss_image_base_dmd2: technical report
2
+
3
+ An 8-step student of MOSS-Image-Base, trained by distribution matching in the style of
4
+ DMD2 (Yin et al., *Improved Distribution Matching Distillation for Fast Image
5
+ Synthesis*, 2024), without its GAN term. The student is a LoRA on the frozen
6
+ transformer; a second LoRA on the same frozen transformer learns the student's own
7
+ output distribution. No image enters a loss: training images only supply caption
8
+ texts and aspect ratios.
9
+
10
+ ## Notation
11
+
12
+ Flow convention: `x_t = (1 - t) x0 + t eps`, with `t = 1` pure noise, `eps` standard
13
+ Gaussian noise.
14
+
15
+ | Symbol | Meaning |
16
+ | --- | --- |
17
+ | `S(x, t)` | Student: the frozen transformer plus the student LoRA, a velocity |
18
+ | `f(x, τ)` | Fake score: the frozen transformer plus the fake LoRA, a velocity |
19
+ | `u(x, τ)` | The frozen teacher on the prompt |
20
+ | `n(x, τ)` | The frozen teacher on the empty prompt with block 22 skipped |
21
+ | `v_real` | The guided teacher, `n + 3 (u - n)` |
22
+ | `x̂` | A clean estimate from the student |
23
+ | `τ` | The noise level at which teacher and fake are compared |
24
+ | `t_r` | Time of sampler step `r`, for `r` in 0..7 |
25
+
26
+ ## Sampler
27
+
28
+ ```
29
+ x = eps # t_0 = 1
30
+ for r in 0..7:
31
+ x̂ = x - t_r S(x, t_r)
32
+ if r < 7:
33
+ x = (1 - t_{r+1}) x̂ + t_{r+1} eps_r # fresh eps_r every step
34
+ output x̂
35
+ ```
36
+
37
+ The times are 8 evenly spaced points 1, 7/8, ..., 1/8 under a log-SNR shift of -3,
38
+ `t_r = s u / (1 + (s - 1) u)` with `u = 1 - r/8` and `s = e^1.5 ≈ 4.4817`: about 1,
39
+ .969, .931, .882, .818, .729, .599, .39. Training uses exactly this chain, and
40
+ sampling should too.
41
+
42
+ ## Objective
43
+
44
+ **Student update.** Run the chain without gradient up to step `r`, take step `r` with
45
+ gradient to get `x̂`, draw a noise level `τ`, noise `x_τ = (1 - τ) x̂ + τ eps`, and
46
+ evaluate `u`, `n` and `f` at `x_τ` without gradient. The force on `x̂` and its loss are
47
+
48
+ ```
49
+ g = τ (v_real - f(x_τ, τ)) / N
50
+ N = mean | x̂ - (x_τ - τ v_real) | # per image, floored at 1e-6
51
+ L = 0.5 mean (x̂ - stop_grad(x̂ - g))²
52
+ ```
53
+
54
+ so the gradient of `L` with respect to `x̂` is `g` (up to the element count): it moves
55
+ the student's output towards the guided teacher's clean estimate and away from the
56
+ fake's. `N`, DMD2's normaliser, reads the same guided teacher as the force.
57
+
58
+ **Fake update.** Run the student's chain without gradient to the output `x̂` of step
59
+ `r`, draw `τ`, noise with fresh `eps`, and train the fake LoRA by flow matching:
60
+ `mean (f(x_τ, τ) - (eps - x̂))²`. The fake sees only the noise level, not the step.
61
+
62
+ **Schedule of updates.** 200 fake-only updates first, then cycles of 5 fake updates
63
+ and 1 student update. A cycle trains one step `r`, shared by its fake and student
64
+ updates, and the next cycle takes the next step, 0 to 7 and round again. So the fake
65
+ has just trained on the outputs of the step it then judges.
66
+
67
+ **Noise levels.** Each update draws one `τ` per image, stratified across the update's
68
+ images: with `B` images, `q_i = (π(i) + ξ_i) / B` for a random permutation `π` and
69
+ uniform `ξ_i`, so each slice of the distribution gets exactly one image.
70
+
71
+ The student's `τ` is uniform and the fake's Beta(2, 2). Both are then shifted by a
72
+ log-SNR shift of -2.2 (`τ' = sigmoid(logit τ + 1.1)`) and not clamped.
73
+
74
+ Under that shift, the fake's Beta(2, 2) has a logit mean of about 1.10 and a standard
75
+ deviation of about 1.14. This is why it was chosen: it sits close to the noise levels
76
+ MOSS-Image itself was pretrained on, a truncated logit-normal with, in this
77
+ convention, a logit mean of 1.09 and a standard deviation of 0.95
78
+ ([MOSS-Image technical report](https://github.com/OpenMOSS/MOSS-Image/blob/main/paper/MOSS_Image_Technical_Report.pdf)).
79
+ The distillation does not use that logit-normal itself.
80
+
81
+ ## Recipe
82
+
83
+ | Setting | Value |
84
+ | --- | --- |
85
+ | Teacher | MOSS-Image-Base, frozen (bf16); text encoder and VAE frozen |
86
+ | Real score | CFG 3, the negative branch being the empty prompt with block 22 skipped, at every noise level |
87
+ | Student and fake | Two LoRAs on the same frozen transformer, each rank 128, alpha 128 (scale 1), A Kaiming-uniform and B zero at initialization, no dropout |
88
+ | Adapted modules | `q`, `k`, `v`, `o`, `ffn_up_gate` and `ffn_down` of all 42 transformer blocks, the time MLP's two linears, the shared block modulation and the head modulation; `ffn_up_gate` carries one rank-128 pair for each of its up and gate halves; the text embedder is not adapted |
89
+ | Optimizer | AdamW for each LoRA, β = (0.9, 0.95), ε 1e-8, no weight decay |
90
+ | Updates | 4,200: 200 fake-only, then 666 cycles of 5 fake and 1 student update (3,534 fake and 666 student updates in all) |
91
+ | Learning rate | 1e-5 for both LoRAs, no warmup: constant for the first 3,300 updates (516 student updates), then a 1 − √ decay to 0 over the last 900 (150 student updates). Both LoRAs follow the same decay, counted in updates of either kind |
92
+ | Gradient clipping | Global norm 10 |
93
+ | Precision | bf16 autocast for the forward and backward passes, TF32 for float32 matrix products; the LoRA factors, AdamW states and EMAs are float32 (the factors are cast to bf16 in the matrix product) |
94
+ | EMAs | Two power-function EMAs of the student's factors (Karras et al.), counted in student updates only: γ 6.94 with its decay capped at 0.9999, and γ 16.97 uncapped; the fake's factors have none |
95
+ | Batch | 16 images per update |
96
+ | Resolution | 29 aspect-ratio buckets of about 1 megapixel (1024 × 1024 and 28 others), sides in multiples of 64, aspect ratios from about 1 : 2.1 to 3.75 : 1 |
97
+ | Text | Captions 80% natural language, 20% JSON layouts (MOSS-Image's layout prompt format); prompts up to 2,048 tokens; no caption dropout |
98
+ | Released weights | an EMA of the student's factors (Karras power function, γ 6.94), stored in bfloat16 |
99
+
100
+ ## Design notes
101
+
102
+ Every parameter and setting above was extensively ablated beforehand on an internal
103
+ 2B text-to-image model with a similar architecture, at lower resolution. These are
104
+ the choices that mattered most, judged by eye on fixed prompts:
105
+
106
+ - **The step order matters most.** Giving the fake and the student independent
107
+ shuffles of the 8 steps produced wrong limbs, duplicated heads and objects within
108
+ a few hundred updates; sharing one step per cycle did not.
109
+ - **Equal learning rates.** A student slower than its fake delayed the early contrast
110
+ overshoot without lowering it.
111
+ - **Time and modulation layers in the LoRAs.** Adapting the time MLP and the
112
+ modulations brought the output close to the teacher's contrast much sooner.
113
+ - **Stratified noise levels** gave a lower contrast peak than independent draws.
114
+ - **Beta(2, 2) for the fake** gave cleaner structure than a uniform draw.
115
+ - **No GAN and no regression loss**: the distribution-matching force alone.
images/example_goose.png ADDED

Git LFS Details

  • SHA256: d84839a38a9252f2b8f10a81c861b0c5afd2df88d46ac673b3e6480e8c7b2b86
  • Pointer size: 132 Bytes
  • Size of remote file: 1.93 MB
loras/moss_image_base_dmd2_bf16.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:36714b3fa11ad665cfc52f803fc6bed6ffedcd829700acc5938b0606957b4e5b
3
+ size 1012265104
transformer/adapter_config.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "base_model_name_or_path": "OpenMOSS-Team/MOSS-Image-Base",
3
+ "bias": "none",
4
+ "inference_mode": true,
5
+ "init_lora_weights": true,
6
+ "lora_alpha": 128,
7
+ "lora_dropout": 0.0,
8
+ "peft_type": "LORA",
9
+ "r": 128,
10
+ "rank_pattern": {
11
+ "ffn_up_gate": 256
12
+ },
13
+ "alpha_pattern": {
14
+ "ffn_up_gate": 256
15
+ },
16
+ "target_modules": "(layers\\.\\d+\\.proj\\.(q|k|v|o|ffn_up_gate|ffn_down)|time_embedding\\.(0|2)|shared_block_modulation|video_head_modulation)",
17
+ "task_type": null,
18
+ "use_dora": false,
19
+ "use_rslora": false
20
+ }
transformer/adapter_model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e717a7aa558e8819b50c09e91efa49ba38ef2a44d5747477695cbb4cd72e786a
3
+ size 814086912
workflows/moss_image_base_dmd2_8step.json ADDED
@@ -0,0 +1,516 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "last_node_id": 11,
3
+ "last_link_id": 11,
4
+ "nodes": [
5
+ {
6
+ "id": 1,
7
+ "type": "UNETLoader",
8
+ "pos": [
9
+ 0,
10
+ 0
11
+ ],
12
+ "size": [
13
+ 320,
14
+ 110
15
+ ],
16
+ "flags": {},
17
+ "order": 1,
18
+ "mode": 0,
19
+ "inputs": [],
20
+ "outputs": [
21
+ {
22
+ "name": "MODEL",
23
+ "type": "MODEL",
24
+ "links": [
25
+ 1
26
+ ],
27
+ "slot_index": 0
28
+ }
29
+ ],
30
+ "properties": {
31
+ "Node name for S&R": "UNETLoader"
32
+ },
33
+ "widgets_values": [
34
+ "moss_image_base_bf16.safetensors",
35
+ "default"
36
+ ]
37
+ },
38
+ {
39
+ "id": 2,
40
+ "type": "LoraLoaderBypassModelOnly",
41
+ "pos": [
42
+ 360,
43
+ 0
44
+ ],
45
+ "size": [
46
+ 320,
47
+ 110
48
+ ],
49
+ "flags": {},
50
+ "order": 2,
51
+ "mode": 0,
52
+ "inputs": [
53
+ {
54
+ "name": "model",
55
+ "type": "MODEL",
56
+ "link": 1
57
+ }
58
+ ],
59
+ "outputs": [
60
+ {
61
+ "name": "MODEL",
62
+ "type": "MODEL",
63
+ "links": [
64
+ 2
65
+ ],
66
+ "slot_index": 0
67
+ }
68
+ ],
69
+ "properties": {
70
+ "Node name for S&R": "LoraLoaderBypassModelOnly"
71
+ },
72
+ "widgets_values": [
73
+ "moss_image_base_dmd2_bf16.safetensors",
74
+ 1.0
75
+ ]
76
+ },
77
+ {
78
+ "id": 3,
79
+ "type": "ModelSamplingAuraFlow",
80
+ "pos": [
81
+ 720,
82
+ 0
83
+ ],
84
+ "size": [
85
+ 320,
86
+ 110
87
+ ],
88
+ "flags": {},
89
+ "order": 3,
90
+ "mode": 0,
91
+ "inputs": [
92
+ {
93
+ "name": "model",
94
+ "type": "MODEL",
95
+ "link": 2
96
+ }
97
+ ],
98
+ "outputs": [
99
+ {
100
+ "name": "MODEL",
101
+ "type": "MODEL",
102
+ "links": [
103
+ 3
104
+ ],
105
+ "slot_index": 0
106
+ }
107
+ ],
108
+ "properties": {
109
+ "Node name for S&R": "ModelSamplingAuraFlow"
110
+ },
111
+ "widgets_values": [
112
+ 4.4817
113
+ ],
114
+ "title": "Flow shift e^1.5 (as trained)"
115
+ },
116
+ {
117
+ "id": 4,
118
+ "type": "CLIPLoader",
119
+ "pos": [
120
+ 0,
121
+ 180
122
+ ],
123
+ "size": [
124
+ 320,
125
+ 110
126
+ ],
127
+ "flags": {},
128
+ "order": 4,
129
+ "mode": 0,
130
+ "inputs": [],
131
+ "outputs": [
132
+ {
133
+ "name": "CLIP",
134
+ "type": "CLIP",
135
+ "links": [
136
+ 4
137
+ ],
138
+ "slot_index": 0
139
+ }
140
+ ],
141
+ "properties": {
142
+ "Node name for S&R": "CLIPLoader"
143
+ },
144
+ "widgets_values": [
145
+ "qwen3.5_9b_bf16.safetensors",
146
+ "moss_image",
147
+ "default"
148
+ ]
149
+ },
150
+ {
151
+ "id": 5,
152
+ "type": "VAELoader",
153
+ "pos": [
154
+ 0,
155
+ 360
156
+ ],
157
+ "size": [
158
+ 320,
159
+ 110
160
+ ],
161
+ "flags": {},
162
+ "order": 5,
163
+ "mode": 0,
164
+ "inputs": [],
165
+ "outputs": [
166
+ {
167
+ "name": "VAE",
168
+ "type": "VAE",
169
+ "links": [
170
+ 10
171
+ ],
172
+ "slot_index": 0
173
+ }
174
+ ],
175
+ "properties": {
176
+ "Node name for S&R": "VAELoader"
177
+ },
178
+ "widgets_values": [
179
+ "flux2-vae_small.safetensors"
180
+ ]
181
+ },
182
+ {
183
+ "id": 6,
184
+ "type": "CLIPTextEncode",
185
+ "pos": [
186
+ 360,
187
+ 180
188
+ ],
189
+ "size": [
190
+ 320,
191
+ 110
192
+ ],
193
+ "flags": {},
194
+ "order": 6,
195
+ "mode": 0,
196
+ "inputs": [
197
+ {
198
+ "name": "clip",
199
+ "type": "CLIP",
200
+ "link": 4
201
+ }
202
+ ],
203
+ "outputs": [
204
+ {
205
+ "name": "CONDITIONING",
206
+ "type": "CONDITIONING",
207
+ "links": [
208
+ 5,
209
+ 6
210
+ ],
211
+ "slot_index": 0
212
+ }
213
+ ],
214
+ "properties": {
215
+ "Node name for S&R": "CLIPTextEncode"
216
+ },
217
+ "widgets_values": [
218
+ "A large brown and white goose resting on dry grass at the edge of a rocky pond, its patterned wing feathers folded neatly and its long neck curved as it looks toward the water. The sun casts warm light across the bird and the rough stone bank behind it."
219
+ ]
220
+ },
221
+ {
222
+ "id": 7,
223
+ "type": "ConditioningZeroOut",
224
+ "pos": [
225
+ 720,
226
+ 180
227
+ ],
228
+ "size": [
229
+ 320,
230
+ 110
231
+ ],
232
+ "flags": {},
233
+ "order": 7,
234
+ "mode": 0,
235
+ "inputs": [
236
+ {
237
+ "name": "conditioning",
238
+ "type": "CONDITIONING",
239
+ "link": 6
240
+ }
241
+ ],
242
+ "outputs": [
243
+ {
244
+ "name": "CONDITIONING",
245
+ "type": "CONDITIONING",
246
+ "links": [
247
+ 7
248
+ ],
249
+ "slot_index": 0
250
+ }
251
+ ],
252
+ "properties": {
253
+ "Node name for S&R": "ConditioningZeroOut"
254
+ },
255
+ "widgets_values": [],
256
+ "title": "Unused at CFG 1"
257
+ },
258
+ {
259
+ "id": 8,
260
+ "type": "EmptyFlux2LatentImage",
261
+ "pos": [
262
+ 360,
263
+ 360
264
+ ],
265
+ "size": [
266
+ 320,
267
+ 110
268
+ ],
269
+ "flags": {},
270
+ "order": 8,
271
+ "mode": 0,
272
+ "inputs": [],
273
+ "outputs": [
274
+ {
275
+ "name": "LATENT",
276
+ "type": "LATENT",
277
+ "links": [
278
+ 8
279
+ ],
280
+ "slot_index": 0
281
+ }
282
+ ],
283
+ "properties": {
284
+ "Node name for S&R": "EmptyFlux2LatentImage"
285
+ },
286
+ "widgets_values": [
287
+ 1152,
288
+ 896,
289
+ 1
290
+ ]
291
+ },
292
+ {
293
+ "id": 9,
294
+ "type": "KSampler",
295
+ "pos": [
296
+ 1080,
297
+ 0
298
+ ],
299
+ "size": [
300
+ 320,
301
+ 110
302
+ ],
303
+ "flags": {},
304
+ "order": 9,
305
+ "mode": 0,
306
+ "inputs": [
307
+ {
308
+ "name": "model",
309
+ "type": "MODEL",
310
+ "link": 3
311
+ },
312
+ {
313
+ "name": "positive",
314
+ "type": "CONDITIONING",
315
+ "link": 5
316
+ },
317
+ {
318
+ "name": "negative",
319
+ "type": "CONDITIONING",
320
+ "link": 7
321
+ },
322
+ {
323
+ "name": "latent_image",
324
+ "type": "LATENT",
325
+ "link": 8
326
+ }
327
+ ],
328
+ "outputs": [
329
+ {
330
+ "name": "LATENT",
331
+ "type": "LATENT",
332
+ "links": [
333
+ 9
334
+ ],
335
+ "slot_index": 0
336
+ }
337
+ ],
338
+ "properties": {
339
+ "Node name for S&R": "KSampler"
340
+ },
341
+ "widgets_values": [
342
+ 42,
343
+ "fixed",
344
+ 8,
345
+ 1.0,
346
+ "lcm",
347
+ "simple",
348
+ 1.0
349
+ ]
350
+ },
351
+ {
352
+ "id": 10,
353
+ "type": "VAEDecode",
354
+ "pos": [
355
+ 1440,
356
+ 0
357
+ ],
358
+ "size": [
359
+ 320,
360
+ 110
361
+ ],
362
+ "flags": {},
363
+ "order": 10,
364
+ "mode": 0,
365
+ "inputs": [
366
+ {
367
+ "name": "samples",
368
+ "type": "LATENT",
369
+ "link": 9
370
+ },
371
+ {
372
+ "name": "vae",
373
+ "type": "VAE",
374
+ "link": 10
375
+ }
376
+ ],
377
+ "outputs": [
378
+ {
379
+ "name": "IMAGE",
380
+ "type": "IMAGE",
381
+ "links": [
382
+ 11
383
+ ],
384
+ "slot_index": 0
385
+ }
386
+ ],
387
+ "properties": {
388
+ "Node name for S&R": "VAEDecode"
389
+ },
390
+ "widgets_values": []
391
+ },
392
+ {
393
+ "id": 11,
394
+ "type": "SaveImage",
395
+ "pos": [
396
+ 1440,
397
+ 180
398
+ ],
399
+ "size": [
400
+ 320,
401
+ 110
402
+ ],
403
+ "flags": {},
404
+ "order": 11,
405
+ "mode": 0,
406
+ "inputs": [
407
+ {
408
+ "name": "images",
409
+ "type": "IMAGE",
410
+ "link": 11
411
+ }
412
+ ],
413
+ "outputs": [],
414
+ "properties": {
415
+ "Node name for S&R": "SaveImage"
416
+ },
417
+ "widgets_values": [
418
+ "moss_image_base_dmd2"
419
+ ]
420
+ }
421
+ ],
422
+ "links": [
423
+ [
424
+ 1,
425
+ 1,
426
+ 0,
427
+ 2,
428
+ 0,
429
+ "MODEL"
430
+ ],
431
+ [
432
+ 2,
433
+ 2,
434
+ 0,
435
+ 3,
436
+ 0,
437
+ "MODEL"
438
+ ],
439
+ [
440
+ 3,
441
+ 3,
442
+ 0,
443
+ 9,
444
+ 0,
445
+ "MODEL"
446
+ ],
447
+ [
448
+ 4,
449
+ 4,
450
+ 0,
451
+ 6,
452
+ 0,
453
+ "CLIP"
454
+ ],
455
+ [
456
+ 5,
457
+ 6,
458
+ 0,
459
+ 9,
460
+ 1,
461
+ "CONDITIONING"
462
+ ],
463
+ [
464
+ 6,
465
+ 6,
466
+ 0,
467
+ 7,
468
+ 0,
469
+ "CONDITIONING"
470
+ ],
471
+ [
472
+ 7,
473
+ 7,
474
+ 0,
475
+ 9,
476
+ 2,
477
+ "CONDITIONING"
478
+ ],
479
+ [
480
+ 8,
481
+ 8,
482
+ 0,
483
+ 9,
484
+ 3,
485
+ "LATENT"
486
+ ],
487
+ [
488
+ 9,
489
+ 9,
490
+ 0,
491
+ 10,
492
+ 0,
493
+ "LATENT"
494
+ ],
495
+ [
496
+ 10,
497
+ 5,
498
+ 0,
499
+ 10,
500
+ 1,
501
+ "VAE"
502
+ ],
503
+ [
504
+ 11,
505
+ 10,
506
+ 0,
507
+ 11,
508
+ 0,
509
+ "IMAGE"
510
+ ]
511
+ ],
512
+ "groups": [],
513
+ "config": {},
514
+ "extra": {},
515
+ "version": 0.4
516
+ }
workflows/moss_image_base_dmd2_8step_api.json ADDED
@@ -0,0 +1,121 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "1": {
3
+ "class_type": "UNETLoader",
4
+ "inputs": {
5
+ "unet_name": "moss_image_base_bf16.safetensors",
6
+ "weight_dtype": "default"
7
+ }
8
+ },
9
+ "2": {
10
+ "class_type": "LoraLoaderBypassModelOnly",
11
+ "inputs": {
12
+ "model": [
13
+ "1",
14
+ 0
15
+ ],
16
+ "lora_name": "moss_image_base_dmd2_bf16.safetensors",
17
+ "strength_model": 1.0
18
+ }
19
+ },
20
+ "3": {
21
+ "class_type": "ModelSamplingAuraFlow",
22
+ "inputs": {
23
+ "model": [
24
+ "2",
25
+ 0
26
+ ],
27
+ "shift": 4.4817
28
+ }
29
+ },
30
+ "4": {
31
+ "class_type": "CLIPLoader",
32
+ "inputs": {
33
+ "clip_name": "qwen3.5_9b_bf16.safetensors",
34
+ "type": "moss_image",
35
+ "device": "default"
36
+ }
37
+ },
38
+ "5": {
39
+ "class_type": "VAELoader",
40
+ "inputs": {
41
+ "vae_name": "flux2-vae_small.safetensors"
42
+ }
43
+ },
44
+ "6": {
45
+ "class_type": "CLIPTextEncode",
46
+ "inputs": {
47
+ "clip": [
48
+ "4",
49
+ 0
50
+ ],
51
+ "text": "A large brown and white goose resting on dry grass at the edge of a rocky pond, its patterned wing feathers folded neatly and its long neck curved as it looks toward the water. The sun casts warm light across the bird and the rough stone bank behind it."
52
+ }
53
+ },
54
+ "7": {
55
+ "class_type": "ConditioningZeroOut",
56
+ "inputs": {
57
+ "conditioning": [
58
+ "6",
59
+ 0
60
+ ]
61
+ }
62
+ },
63
+ "8": {
64
+ "class_type": "EmptyFlux2LatentImage",
65
+ "inputs": {
66
+ "width": 1152,
67
+ "height": 896,
68
+ "batch_size": 1
69
+ }
70
+ },
71
+ "9": {
72
+ "class_type": "KSampler",
73
+ "inputs": {
74
+ "model": [
75
+ "3",
76
+ 0
77
+ ],
78
+ "positive": [
79
+ "6",
80
+ 0
81
+ ],
82
+ "negative": [
83
+ "7",
84
+ 0
85
+ ],
86
+ "latent_image": [
87
+ "8",
88
+ 0
89
+ ],
90
+ "seed": 42,
91
+ "steps": 8,
92
+ "cfg": 1.0,
93
+ "sampler_name": "lcm",
94
+ "scheduler": "simple",
95
+ "denoise": 1.0
96
+ }
97
+ },
98
+ "10": {
99
+ "class_type": "VAEDecode",
100
+ "inputs": {
101
+ "samples": [
102
+ "9",
103
+ 0
104
+ ],
105
+ "vae": [
106
+ "5",
107
+ 0
108
+ ]
109
+ }
110
+ },
111
+ "11": {
112
+ "class_type": "SaveImage",
113
+ "inputs": {
114
+ "images": [
115
+ "10",
116
+ 0
117
+ ],
118
+ "filename_prefix": "moss_image_base_dmd2"
119
+ }
120
+ }
121
+ }